Automatic Multiscale Image Segmentation
Speaker: Narendra Ahuja , University of Illinois at Urbana-ChampaignContact:
Date: April 26 2004
Time: 11:00AM to 12:00PM
Host: W. Eric L. Grimson, CSAIL
Greg Shakhnarovich, firstname.lastname@example.orgRelevant URL:
This talk is in two parts. First, we present an overview of some recent and ongoing projects in our laboratory. In the second part, we focus on our work on automatic segmentation of image data, aimed at extracting image regions at all scales that happen to be present in a given image. Our approach treats the image as a grid of charged particles and elicits the segmentation in terms of the structures formed by inter-particle forces. The result is not a single segmentation but a nesting of all pertinent segmentations, represented as a segmentation tree. The tree captures all regions, having different degrees of smoothness, detected irrespective of their shapes, sizes, topological relationships and contrasts.
As a method for independent, statistical characterization and estimation of the above segmentation, we present our explorations into the use of Parzen Windows for modeling the image variation. The validity of such a model is shown to follow naturally from the elementary gestalt laws of vicinity, similarity and continuity of direction. We present consistency results for Parzen Window estimators, and describe a plug-in estimator for the scale of the window kernels which is shown to be asymptotically optimal. We evaluate the segmentation algorithms developed on real data.
See other events that are part of MIT Machine Vision Colloquium Spring 2004
See other events happening in April 2004